HosseinAdeli/BLT_recurrent_models

BLT recurrent models: run 1

This collection contains 75 best checkpoints from run 1, approximately 10.38 GiB, with training parameters and validation logs where available. The Apache-2.0 metadata is retained from this repository's existing model card.

Training and architecture code

Contents

Objective	Dataset	Models
classification	imagenet	24
classification	imagenet_vggface2	24
classification	vggface2	24
lejepa	imagenet	2
visreg	imagenet	1

Files retain their original paths: model_family/<objective>/<dataset>/<model>/run_1/checkpoint.pth. See run_1_manifest.json for the complete inventory. Only each run's top-level best checkpoint is included in this release; nested epoch snapshots and other run numbers are not added. Previously published files are retained.

These are original PyTorch training checkpoints, including optimizer/scheduler state when saved, rather than Transformers models or weights-only exports. Training parameters differ between experiments: use the saved checkpoint's args and accompanying params.txt rather than today's training defaults. Validation logs record improvements and may use historical metric calculations; reported values should not be assumed directly comparable across experiments.

Download one model

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="HosseinAdeli/BLT_recurrent_models",
    filename="model_family/classification/imagenet/blt_bl/run_1/checkpoint.pth",
)

Download all run_1 checkpoints

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="HosseinAdeli/BLT_recurrent_models",
    local_dir="BLT_recurrent_models_weights",
    allow_patterns=["model_family/*/*/*/run_1/checkpoint.pth",
                    "model_family/*/*/*/run_1/params.txt",
                    "model_family/*/*/*/run_1/val_results.txt",
                    "run_1_manifest.json", "README.md"],
)

Load with the training code

Install the training repository's dependencies and run this from its root after downloading a checkpoint. Use a code version supporting the checkpoint's objective. Only load pickle-based checkpoints from a source you trust.

import torch
from models.build_model import build_model

checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
args = checkpoint["args"]
# Older supervised checkpoints predate the projection-head argument.
if not hasattr(args, "proj_dim"):
    args.proj_dim = 512
if not hasattr(args, "objective"):
    args.objective = "classification"
model = build_model(args, pretrained=False)
model.load_state_dict(checkpoint["model"], strict=True)
model.eval()

Use the corresponding dataset preprocessing from the training code. SSL models also contain a trained online classification readout. No training images are included in this repository.